InstructLab
An open-source way to add your own facts and skills to an existing large language model, instead of looking those facts up at question time the way retrieval does. IBM Research and Red Hat built it, an IBM Power Developer post ran it on a Power 10 partition with no GPU, and IBM says it turned a Granite code model into an IBM Z modernization expert in one week.
What InstructLab does
A large language model, or LLM, learns from general text. It does not know your company rules, your part numbers, or your own programs. Retrieval is one answer to that: look the facts up at question time and hand them to the model inside the prompt. InstructLab is a different answer. It changes the model. IBM Research describes it as lowering the cost of fine-tuning large language models by allowing people to collaboratively add new knowledge and skills to any model.
Think about a new hire. Retrieval is handing them a binder to flip through on every call. InstructLab is sending them to class until they know the material. The binder is quicker to put together. The class makes a faster worker.
The name holds the method. LAB stands for Large-scale Alignment for chatBots. The paper, titled "LAB: Large-Scale Alignment for ChatBots", went up on arXiv on March 2, 2024 and names the MIT-IBM Watson AI Lab and IBM Research on its front page. arXiv is the open site where research papers are posted before or beside formal publication.
How does the tuning work?
The three steps IBM lists
- Write down what you want it to learnYou add examples to a taxonomy, a tree of folders holding YAML files. YAML is a plain text format for structured notes. IBM Research calls the taxonomy a hierarchical map of what models tuned on InstructLab data have learned to date. Each entry is either knowledge, meaning facts, data or references, or a skill, meaning something you want the model to do.
- Let it write the rest of the training dataSynthetic data generation grows your small seed set into a large training set. A bigger teacher model writes the extra examples and a student model learns from them. The LAB paper says this reaches competitive results without human-annotated data and without a proprietary model such as GPT-4 doing the writing.
- Tune in phases, knowledge before skillsAlignment tuning runs knowledge tuning first and skill tuning second. IBM Research says the regimen lets new information be assimilated into the model during alignment, without causing the model to overwrite what it previously learned.
Steps as listed on the IBM topic page "What is InstructLab?". The wording about not overwriting comes from the IBM Research post "A new way to collaboratively customize LLMs", dated June 7, 2024. Both read October 25, 2026.
Overwriting is the usual risk when you train a finished model on new material. It has a name: catastrophic forgetting. The model picks up your facts and quietly loses something it already knew. Avoiding it is the claim the LAB paper rests on.
Can an IBM i shop run it?
Not under IBM i itself. InstructLab is a Linux program. The useful part is that it runs on the same Power hardware, in a partition of its own.
- An IBM Power Developer post by Marvin Giessing, dated May 24, 2024, ran it on a Power 10 LPAR. An LPAR, or logical partition, is a slice of one Power server running its own operating system.
- Red Hat Enterprise Linux 8.x or 9.x, with 16 virtual processors and 32 GB of memory recommended for best performance.
- No GPU. A GPU, or graphics processing unit, is the add-in card most AI training leans on.
- It ran inside a container started with podman, with the Granite 7b model downloaded separately and mounted into the container.
- Simultaneous multithreading set to 2 with the ppc64_cpu command before the run.
- The commands shown were ilab init to set it up, ilab chat to talk to the model, and ilab serve to publish it as an endpoint other programs can call.
The speed came from the chip, not from a card. That post says the work runs on llama.cpp through its llama-cpp-python bindings, with hardware acceleration for vector operations through VSX and Altivec and for matrix operations through the Matrix Math Accelerator, or MMA, by way of the OpenBLAS library. MMA is the unit IBM builds into the Power 10 and Power 11 chips, so the matrix work has somewhere to go without a GPU.
We found no IBM documentation of InstructLab running under IBM i or inside PASE for i, checked October 25, 2026. Plan on a Linux partition or a separate machine until IBM says otherwise.
Which models come already tuned?
You do not have to tune anything to get something out of this. The IBM announcement for InstructLab support in watsonx.ai lists language models that were already tuned with the method, naming granite-7b-lab, merlinite-7b, granite-20b-multilingual and granite-13b-chat-v2. It also lists open code models you can reach through watsonx.ai.
The open code models IBM lists with InstructLab support in watsonx.ai, by size in billions of parameters
A parameter is one of the adjustable numbers inside a model. More parameters usually means a bigger machine to run it. granite-20b is marked because 20 billion is the size IBM used in the mainframe test below. Read from the IBM announcement "Democratizing Large Language Model development with InstructLab support in watsonx.ai" on October 25, 2026.
IBM says those models arrive either through a commercial Red Hat Enterprise Linux AI subscription from Red Hat, or directly within watsonx.ai. IBM also says Red Hat merges community contributions on a weekly cadence, so the shared taxonomy keeps moving without you doing the tuning.
What happened when IBM tried it on old code?
InstructLab in its first months, by dated source
- March 2, 2024The LAB paper goes up on arXiv, describing a taxonomy-guided synthetic data generation process and a multi-phase tuning framework.
- May 24, 2024An IBM Power Developer post shows InstructLab running on a Power 10 logical partition under Red Hat Enterprise Linux, with no GPU.
- June 7, 2024IBM Research publishes "A new way to collaboratively customize LLMs", reporting the IBM Z code result below.
Dates taken from each source on October 25, 2026.
That last one is the reason this term belongs on an IBM i site. IBM Research says it took an IBM 20B Granite code model and turned it into an expert at modernizing software written for IBM Z mainframes. It says the work took one week and reached a code generation score of 97 percent.
IBM Z is not IBM i. They are different machines, and the old code on Z is mostly COBOL rather than RPG. RPG is the language most IBM i business applications are written in, and COBOL runs on both. What carries over is the shape of the job: take a general code model, feed it one narrow old language, and get a specialist back.
The 97 percent is an IBM number from an IBM test, and IBM has not published the same test for RPG. Read it as a sign the method works on an old language, not as a measurement of anything on your system.
Related Terms
IBM Granite
IBM's family of open-weight foundation models, released under the Apache 2.0 license, that underpins watsonx.ai and serves as the base model fine-tuned for RPG and IBM i in watsonx Code Assistant for i and IBM Bob.
Large Language Model (LLM)
A machine learning model trained on large volumes of text, used as the underlying technology behind generative AI tools including IBM Bob and general-purpose assistants like ChatGPT and Claude.
Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation hands a language model passages pulled from your own data, so it answers from your content. On IBM i, retrieval is the hard half.
Matrix Math Acceleration (MMA)
An on-chip acceleration feature in IBM Power10 and Power11 processors that speeds up the matrix math operations used in AI inferencing, without requiring a separate GPU.
IBM Power 10
IBM's Power processor generation that preceded Power11, announced August 17, 2020 and shipping from September 2021, the first Power chip to embed Matrix Math Acceleration into every core for on-chip AI inferencing without a separate GPU. IBM began winding down new Power10 sales in February 2026.
Red Hat OpenShift AI
Red Hat's add-on platform for serving and governing AI models on a Red Hat OpenShift cluster, supported self-managed on ppc64le, the IBM Power architecture that also runs IBM i. It lets a Power shop host open weight models on its own hardware instead of calling an outside service.
watsonx.ai
The studio component of watsonx for building, tuning, and deploying foundation models and generative AI applications, including retrieval-augmented generation (RAG) against enterprise data.
COBOL
A long-established business programming language still running production logic on IBM i and other legacy enterprise platforms alongside RPG.